Person Re-identification Algorithm Based on Logistic Regression Loss Function

Authors:
Open Access
Publisher website
Google Scholar
Abstract
The feature description of pedestrians is the key to the task of person re-identification. The deep feature extracted from convolutional neural networks proved to have good robustness. Deep feature learning based on convolutional neural network also achieved good performance in person re-identification tasks. However, the feature discriminant is not strong under the constraint of traditional classification loss function, which seriously restricts the person re-identification performance. In order to enhance the discrimination of features, this paper uses the idea of logistic regression to associate the distance between samples and the similarity probability and proposes a logistic regression loss function to constrain the training of the network. The logistic regression loss function proposed in this paper can constrain the network to make the distance between similar samples small, and the distance between different classes of samples large. Secondly, by introducing the positive sample distance threshold, this paper avoids that the network constrains the same sample with some very different samples from being too small, while the distance constraint to other samples fails. Finally, the proposed logic regression loss function is embedded into the Siamese convolutional neural network for end-to-end person re-identification. This paper validates the effectiveness of the algorithm on the mainstream databases of Market1501 and CUHK03.